The research on music information retrieval has recently experienced a tremendous growth, due to the exploding number of applications on entertainment or educational areas, for example. In particular, tempo induction systems play a key role for diverse applications such as automatic music recommendation, automatic synchronization or advanced audio editing. Several approaches to the problem of inducting the “perceptual tempo” in musical signals have been recently developed. There are still some unsolved problems and a satisfactory system working on a broad scope of music genres has not been presented yet. In this work the state-of-the-art of tempo induction systems is analyzed, the problem decomposed and key solutions for each stage identified. The use of multidisciplinary knowledge from different areas is used, such as digital signal processing, computer science and machine learning, psychoacoustics, psychology and music. Low level descriptors are extracted out of the raw audio signal. Musical events of different kinds, main responsible of the rhythm percept, are detected. Periodicities underlying the music are extracted and a probabilistic model used to select the most likely perceptual tempo amongst them. A Hidden Markov model is implemented encoding psychological and musical knowledge of higher level of abstraction. Several evaluations are conducted on various music databases and the proposed system is compared to recently presented algorithms.